acceptodds
Under review as a conference paper at ICLR 2027

QUEST: Quantile-Conditioned Selection of Historical Context for Probabilistic Time Series Forecasting

Abstract

Probabilistic time series forecasting (PTSF) plays an important role in decision-making across a wide range of real-world applications, including economics, transportation, energy, and AIOps. Recent direct prediction methods offer an efficient way to produce probabilistic forecasts without assuming a predefined distribution family or relying on iterative sampling. However, existing methods often construct a shared historical context for different parts of the predictive distribution, while the relevance of historical information may vary across quantile levels. This motivates historical context construction that explicitly accounts for the target probability level. In this paper, we propose QUEST, a probabilistic time series forecasting model based on quantile-conditioned historical selection. Specifically, QUEST divides the observed history into patches and encodes them into shared local representations. For each target quantile, a corresponding query interacts with these representations to determine patch weights conditioned on both the observed history and the target probability level, yielding a quantile-specific predictive context. A shared decoder then combines each context with its corresponding query to predict the target quantile over the entire forecast horizon, enabling multiple quantile forecasts in a single forward pass. Extensive experiments on eight real-world multivariate datasets demonstrate that QUEST achieves state-of-the-art probabilistic forecasting performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Probabilistic Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_t48mOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed to use quantile conditioned context aggregation for probabilistic time series forecasting. Specifically, for each quantile, a query vector is constructed and is used to compute weights to combine different patches in the context, resembling the attention mechanism. On a comparison to 10 other probabilistic forecasting baselines, the proposed method seems outperform all the others on both CRPS (for probabilistic forecasts accuracy) and NMAE (for point forecasts accuracy). The proposed model is also very fast in inference due to its small size.

Critical strengths
  • The paper is easy to read and follow.
  • The proposed idea is simple and seems to working well compared with other baselines.
  • The proposed model is small and fast in inference.
Critical weaknesses
  • The motivation is not clear to me. The author wrote "...the relevance of historical information may vary across quantile levels." Why is this true? Why would a quantile focus more on an earlier path than the later patch? In the extreme case, for a short context window that contains only one patch, one would still expect good predicted quantiles.
  • The importance of the proposed query dependent patch aggregation machanism is not clear to me. I can think of two more variants of the methods: a) same weights for all the queries b) instead of trying to aggregate patches, just concatenate patches as longer input (no aggregation will be needed). The comparison to a highlights the impact of the query dependence; the comparison to b indicates if we should do aggregation at all.
Questions for authors
  • Besides my comments on the paper weakness, can the authors explain how did they ensure the baselines are correctly implemented and used? Why did they choose those 10 baselines not the others?
  • There is a typo in Line 257, what is eqq=1Qeqq = 1^Q?
9 comments
AuthorsOpenReview ↗

Response to Reviewer t48m

We sincerely thank you for your detailed review and for recognizing the simplicity of our approach, the clarity of our paper, and the compact model size and fast inference of QUEST. We hope the following responses address your concerns.


W1: Motivation and the single-patch case

Thank you for pointing out the ambiguity in our motivation. We would like to clarify that different quantiles may benefit from different historical information, regardless of its position within the observed window. To make this motivation clearer, we revised the Abstract (lines 17–20)‾\underline{\text{Abstract (lines 17–20)}} of the revised manuscript.

To further illustrate this, we added visualizations in Section 5.3.1, Figure 3‾\underline{\text{Section 5.3.1, Figure 3}} of the revised manuscript, with additional examples in Appendix B.4, Figures 6–9‾\underline{\text{Appendix B.4, Figures 6–9}}. For example, in Appendix B.4, Figure 6‾\underline{\text{Appendix B.4, Figure 6}}, as the quantile level increases from 0.05 to 0.95, the middle patch's weight increases from 5.2% to 85.9% in one ETTh1 example and decreases from 73.2% to 5.2% in another. These results illustrate that historical selection depends on both the input history and the target quantile.

Dividing historical observations into multiple patches is a common practice in time series forecasting [1–3], enabling compact representations of local historical patterns. In the single-patch case, the aggregation weight is 1, and QUEST can still produce different quantile forecasts through its quantile-conditioned decoder. With multiple patches, QUEST can further select historical information according to different quantile levels.

[1] Nie et al. A Time Series Is Worth 64 Words: Long-term Forecasting with Transformers. ICLR 2023.

[2] Lee et al. Learning to Embed Time Series Patches Independently. ICLR 2024.

[3] Wu et al. Enhancing Time Series Forecasting through Selective Representation Spaces: A Patch Perspective. NeurIPS 2025.


W2: Importance of query-dependent patch aggregation

Thank you for your suggestion. To further validate the effectiveness of query-dependent patch aggregation, we conducted two ablation studies following your recommendation: SharedWeight, which uses shared aggregation weights across quantiles, and SharedContext, which concatenates encoded historical patches. In addition, we conducted a SharedQuery ablation to examine the effect of sharing queries across quantile levels.

The table below reports CRPS / NMAE averaged over four forecasting horizons (lower is better).

VariantETTh1ETTm2ElectricityWeather
SharedContext0.285 / 0.3740.144 / 0.1790.080 / 0.1010.065 / 0.082
SharedQuery0.277 / 0.3650.140 / 0.1750.078 / 0.0980.064 / 0.080
SharedWeight0.274 / 0.3660.138 / 0.1750.077 / 0.0980.063 / 0.080
QUEST0.271 / 0.3620.137 / 0.1730.076 / 0.0970.062 / 0.079

These results support the effectiveness of query-dependent patch aggregation.

The ablation results are presented in Section 5.4.1, Table 3‾\underline{\text{Section 5.4.1, Table 3}} of the revised manuscript, with detailed results for each forecasting horizon provided in Appendix B.1, Table 5‾\underline{\text{Appendix B.1, Table 5}}.


Q1: Baseline selection and implementation

1. Baseline Selection

Thank you for your question. To cover different probabilistic forecasting paradigms, we selected representative baselines based on the discussion in the Introduction. We also added two recent methods, CWFlow and TimePrism, bringing the total to twelve baselines.

2. Implementation and Fair Comparison

To ensure a fair comparison, we used the baseline implementations already available in ProbTS [4] and integrated the publicly available code of additional baselines into the same framework. All methods followed consistent data splitting, preprocessing, and evaluation protocols. The experimental settings are described in Section 5.1 (lines 338–364)‾\underline{\text{Section 5.1 (lines 338–364)}}, and the code repositories are provided in Appendix A.4, Table 4‾\underline{\text{Appendix A.4, Table 4}}.

[4] Zhang et al. ProbTS: Benchmarking Point and Distributional Forecasting across Diverse Prediction Horizons. NeurIPS 2024.


Q2: Typo

Thank you for pointing out the typo. We corrected the notation to {eq}q=1Q\left\lbrace e_q \right\rbrace_{q=1}^{Q} in Section 4.2.1, line 264‾\underline{\text{Section 4.2.1, line 264}} of the revised manuscript, where eqe_q is the query for the qq-th target quantile and QQ is the number of target quantile levels. We also carefully checked the manuscript for similar errors.


We hope that our clarifications and additional experiments address your concerns and provide a clearer understanding of our motivation and the effectiveness of the proposed method. Thank you again for your valuable feedback and consideration.

Reviewer_t48mOpenReview ↗

Response to authors

Thank you for the reply and the extra ablation.

Figure 3 indeed shows that QUEST assign different weights of patches for different queries. Having a look of the new ablation table (Table 3), the shared weights (same weights for all patches) is only slightly worse than QUEST. And the biggest improvement QUEST get is over the SharedContext baseline.

How is SharedContext implemented? Do you still learn key and values for SharedContext or there is no attention module in that setting?

Thank you for your further feedback. Your understanding is correct. In SharedContext, there is no attention module, and the encoded historical patches are directly concatenated to construct a shared historical context. Therefore, no Key or Value projections are learned.

Reviewer_QVVxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes QUEST, a probabilistic forecasting model. The model encodes each quantile level as a query embedding, and computes the query embedding's similarity with embedded historical patches. The similarity score is then used as the weight for computing the weighted sum of the historical context, which is then passed to a final decoder to generate the prediction. Extensive experiments on 8 datasets and with 10 baselines show the competitiveness of the proposed method.

Critical strengths
  • The idea is simple and easy to follow. The writing is clear.
  • The experimental results appear to be strong, the proposed QUEST performs best in most datasets.
  • The inference time of the proposed method is also notably good
Critical weaknesses
  • The central claim is that the relevant of historical context can differ across different quantile level. But I believe this claim is not justified through an ablation study. I would be more convinced if the authors could visualize the weights a_{c,q,m} in equation 10 across different quantile levels, and show that they indeed change cross quantile levels. Additionally, an ablation study which uses the same query embedding for all quantile levels, and compared the performance with QUEST, would strengthen the claim.

  • The pinball loss at specific quantile levels, e.g. at 0.05, 0.5, 0.95, and empirical coverages, e.g. MSIS, are not reported for the experiments.

  • Comparing the model sizes of QUEST and baselines could strengthen the paper. As it is unclear whether the performance gain comes from simply using a larger model.

  • Several related classical papers on quantile-based multi-horizon forecasting are not discussed, including them would strengthen the related work section:

[1] Eisenach, C., Patel, Y., & Madeka, D. (2020). Mqtransformer: Multi-horizon forecasts with context dependent and feedback-aware attention. arXiv preprint arXiv:2009.14799.

[2] Gouttes, A., Rasul, K., Koren, M., Stephan, J., & Naghibi, T. (2021). Probabilistic time series forecasting with implicit quantile networks. arXiv preprint arXiv:2107.03743.

[3] Park, Y., Maddix, D., Aubet, F. X., Kan, K., Gasthaus, J., & Wang, Y. (2022, May). Learning quantile functions without quantile crossing for distribution-free time series forecasting. In International conference on artificial intelligence and statistics (pp. 8127-8150). PMLR.

[4] Kan, K., Aubet, F. X., Januschowski, T., Park, Y., Benidis, K., Ruthotto, L., & Gasthaus, J. (2022, May). Multivariate quantile function forecaster. In International Conference on Artificial Intelligence and Statistics (pp. 10603-10621). PMLR.

Questions for authors

Please see weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer QVVx

Thank you for recognizing the simplicity, clarity, and strong performance of QUEST, and for your constructive suggestions. We hope to address your concerns below.


W1: Evidence for quantile-conditioned historical selection

1. Visualization of Historical Selection

To further illustrate quantile-conditioned historical selection, we added visualizations of patch aggregation weights across 19 quantile levels in Section 5.3.1, Figure 3‾\underline{\text{Section 5.3.1, Figure 3}} of the revised manuscript. On ETTh1, as the quantile level increases from 0.05 to 0.95, the middle patch's weight decreases from 79.3% to 5.7%, while the most recent patch's weight increases from 4.3% to 56.7%. Additional examples in Appendix B.4, Figures 6–9‾\underline{\text{Appendix B.4, Figures 6–9}} further illustrate the quantile-conditioned historical selection behavior of QUEST.

2. SharedQuery Ablation

To further validate the effectiveness of quantile-conditioned historical selection, we conducted the SharedQuery ablation following your suggestion, which used a single query shared across quantile levels. In addition, we conducted two ablations, SharedContext and SharedWeight, to examine patch concatenation and shared aggregation weights.

The table below reports CRPS / NMAE averaged over four forecasting horizons (lower is better).

VariantETTh1ETTm2ElectricityWeather
SharedContext0.285 / 0.3740.144 / 0.1790.080 / 0.1010.065 / 0.082
SharedQuery0.277 / 0.3650.140 / 0.1750.078 / 0.0980.064 / 0.080
SharedWeight0.274 / 0.3660.138 / 0.1750.077 / 0.0980.063 / 0.080
QUEST0.271 / 0.3620.137 / 0.1730.076 / 0.0970.062 / 0.079

These results support the effectiveness of quantile-specific queries for historical selection.

The ablation results are presented in Section 5.4.1, Table 3‾\underline{\text{Section 5.4.1, Table 3}} of the revised manuscript, with detailed results for each forecasting horizon provided in Appendix B.1, Table 5‾\underline{\text{Appendix B.1, Table 5}}.


W2: Pinball loss and prediction interval evaluation

Thank you for your suggestion. To further evaluate quantile-level accuracy and prediction interval quality, we evaluated pinball loss at five quantile levels (0.05, 0.25, 0.50, 0.75, and 0.95) and MSIS for the central 90% prediction interval, comparing QUEST with four baselines across six datasets.

The table below reports pinball loss at five quantile levels and MSIS for the central 90% prediction interval, averaged over four forecasting horizons (lower is better).

DatasetMethod0.050.250.500.750.95MSIS₉₀%
ETTm1OCE-TS0.4660.9601.0090.6940.2328.409
DistPred0.4350.7900.8280.6390.2958.725
PPM0.4310.7560.7480.6330.2718.593
TimePrism0.3900.6810.7380.7050.3478.541
QUEST0.3290.6270.7160.5690.2096.870
WeatherOCE-TS2.1785.2407.2357.2874.6327.203
DistPred2.2465.8418.1347.8104.6585.389
PPM3.5716.8618.4457.9055.2529.127
TimePrism3.6356.1609.50311.0456.9047.307
QUEST1.4664.5566.5906.2803.4784.929

These results support the effectiveness of QUEST in quantile forecasting and prediction interval estimation.

Detailed pinball loss and MSIS results for each forecasting horizon are provided in Appendix B.3, Tables 8–10‾\underline{\text{Appendix B.3, Tables 8–10}} of the revised manuscript.


W3: Comparison of model sizes

Thank you for your suggestion. To address your concern about whether QUEST's performance gains stem from a larger model, we compared its parameter counts with those of three representative baselines: PPM, K²VAE, and DistPred.

The table below reports the number of parameters (in millions) on Weather and Electricity across two forecasting horizons (lower is better).

DatasetHorizonQUESTPPMK²VAEDistPred
Weather961.71‾\underline{1.71}1.771.332.17
Weather7202.352.41‾\underline{2.41}3.037.55
Electricity960.46‾\underline{0.46}0.346.742.17
Electricity7200.78‾\underline{0.78}0.508.457.55

These results show that QUEST maintains a competitive model size across different forecasting settings.


W4: Related Work on Quantile-Based Forecasting

Thank you for your suggestion. To provide a more comprehensive discussion of quantile-based multi-horizon forecasting, we added the four recommended studies and expanded the related work in Section 2.1 (lines 123–127)‾\underline{\text{Section 2.1 (lines 123–127)}} of the revised manuscript.


We hope these additional experiments and revisions address your concerns. Thank you again for your valuable feedback.

Dear Reviewer,

Since the End of the Rebuttal comes soon, we would like to inquire if our response addresses your primary concerns. If it does, we kindly request that you reconsider the score. If you have any additional suggestions, we are more than willing to engage in further discussions and make necessary improvements to the paper. Thanks again for dedicating your time to enhancing our paper!

Kind Regards,

Authors

AuthorsOpenReview ↗

Kindly Request for Reviewer's Feedback

Dear Reviewer QVVx,

We sincerely appreciate your efforts during this tight review timeline and your detailed, constructive comments on our paper.

As the end of the rebuttal period is approaching, we would like to kindly check whether our responses have addressed your primary concerns. Following your suggestions, we have added the corresponding experiments and analyses to the revised paper. If our responses have resolved your concerns, we hope you might consider reassessing the score so that we have the opportunity to present this work at the conference.

We are eager to discuss any remaining concerns. Thank you again for dedicating your time to enhancing our paper!

Kind regards,

Authors

Reviewer_am33OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

QUEST predicts marginal quantiles by encoding historical patches independently and pooling them with probability-level queries. A shared decoder combines each quantile-specific context with its query to forecast the full horizon. The model uses pinball-loss training and sorts quantiles during inference.

Critical strengths
  1. Conditioning historical aggregation on the target quantile is a clear and useful modeling idea. Sharing local representations while learning different pooling weights provides a simple way to accommodate different informational requirements across the predictive distribution.
  2. The evaluation demonstrates a competitive practical trade-off. Improvements are particularly substantial on Weather and Traffic, with horizon-averaged CRPS reductions of 11.8% and 8.0% over the strongest reported baselines. The full horizon results and five-run summaries support the breadth of the evaluation, while the reported 0.82 ms and 1.08 ms inference times on ETTh1 and Weather make the approach attractive for efficient forecasting.
Critical weaknesses
  1. The strong overall results would be more convincing with a direct test of the central selection mechanism. Section 5.3 varies decoder width and patch length, but does not compare against shared historical pooling while retaining the same patch encoder, quantile-conditioned decoder, loss, and comparable parameter budget. This is my principal concern: the current evidence does not distinguish the benefit of quantile-specific historical selection from decoder conditioning or other architectural choices. A focused control would clarify the scientific contribution behind the promising forecasting results.
  2. Some relatively recent baselines should be added to the comparison, such as Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting (Yang et al., ICLR 2026) and From Samples to Scenarios: A New Paradigm for Probabilistic Forecasting (Dai et al., ICLR 2026).
Questions for authors
  1. Could you compare quantile-specific pooling with a shared historical context, retaining the same patch encoder, quantile-conditioned decoder, training objective, and comparable model capacity? How much of the forecasting improvement is attributable to the proposed selection mechanism?
  2. How does QUEST compare with some recent models, as mentioned before?
  3. QUEST is a promising and practical approach. I would be inclined to raise my rating if the core ablation leads to meaningful insights and the additional comparisons support a favorable accuracy-efficiency trade-off. For my rating to be raised, the performance of QUEST does not necessarily need to be SOTA. I would like to see more insight into the potential of this paradigm (history selection for quantile).
3 comments
AuthorsOpenReview ↗

Response to Reviewer am33

We sincerely thank you for your constructive comments and for recognizing quantile-conditioned historical selection as a clear and useful idea, as well as the competitive accuracy–efficiency trade-off of QUEST. We hope the following responses address your concerns.


W1 / Q1: Contribution of quantile-conditioned historical selection

Thank you for your suggestion. To further validate the effectiveness of quantile-conditioned historical selection, we conducted the SharedContext ablation following your recommendation. In addition, we conducted two ablations, SharedQuery and SharedWeight, to further examine the effects of sharing queries and aggregation weight across quantile levels.

The table below reports CRPS / NMAE averaged over four forecasting horizons (lower is better).

VariantETTh1ETTm2ElectricityWeather
SharedContext0.285 / 0.3740.144 / 0.1790.080 / 0.1010.065 / 0.082
SharedQuery0.277 / 0.3650.140 / 0.1750.078 / 0.0980.064 / 0.080
SharedWeight0.274 / 0.3660.138 / 0.1750.077 / 0.0980.063 / 0.080
QUEST0.271 / 0.3620.137 / 0.1730.076 / 0.0970.062 / 0.079

These results support the effectiveness of quantile-conditioned historical selection.

The ablation results are presented in Section 5.4.1, Table 3‾\underline{\text{Section 5.4.1, Table 3}} of the revised manuscript, with detailed results for each forecasting horizon provided in Appendix B.1, Table 5‾\underline{\text{Appendix B.1, Table 5}}.

We also added visualizations in Section 5.3.1, Figure 3‾\underline{\text{Section 5.3.1, Figure 3}}, with additional examples in Appendix B.4, Figures 6–9‾\underline{\text{Appendix B.4, Figures 6–9}}, to illustrate the quantile-conditioned historical selection behavior of QUEST.


W2 / Q2 / Q3: Recent baselines and accuracy–efficiency trade-off

1. Comparison with Recent Baselines

Thank you for recommending these recent methods. Following your suggestion, to better compare QUEST with recent related work, we added CWFlow and TimePrism as additional baselines.

The table below reports CRPS / NMAE averaged over four forecasting horizons (lower is better). Due to excessive time and memory consumption, some results are unavailable and denoted as "-".

DatasetQUEST CRPSTimePrism CRPSCWFlow CRPSQUEST NMAETimePrism NMAECWFlow NMAE
ETTm10.2470.2790.4140.3170.3200.457
ETTm20.1370.1600.3770.1730.1840.453
ETTh10.2710.3080.4660.3620.3860.520
ETTh20.1560.1810.4280.2010.2160.472
Electricity0.0760.077—0.0970.097—
Traffic0.1770.206—0.2210.256—
Weather0.0620.1000.0960.0790.1160.117
ILI0.1060.1600.1400.1330.2220.150

These results further demonstrate the superior forecasting performance of QUEST compared with recent baselines.

The full comparison results are presented in Section 5.2, Table 2‾\underline{\text{Section 5.2, Table 2}} of the revised manuscript, with detailed results for each forecasting horizon provided in Appendix B.2, Tables 6 and 7‾\underline{\text{Appendix B.2, Tables 6 and 7}}.

2. Accuracy–Efficiency Comparison

To further evaluate QUEST's accuracy–efficiency trade-off, we compared QUEST with CWFlow and TimePrism.

The table below reports CRPS, inference time per sample, and peak allocated GPU memory on ETTh1 and Weather, with both the context length and forecast horizon set to 96 and a batch size of one (lower is better).

ModelETTh1 CRPSTime (ms)Memory (MB)Weather CRPSTime (ms)Memory (MB)
QUEST0.2370.8215.40‾\underline{15.40}0.0571.08‾\underline{1.08}21.50‾\underline{21.50}
TimePrism0.262‾\underline{0.262}0.85‾\underline{0.85}12.200.084‾\underline{0.084}0.8215.20
CWFlow0.344795.2583.700.093791.6891.10

These results demonstrate that QUEST achieves a favorable accuracy–efficiency trade-off, combining strong forecasting performance with low inference latency and GPU memory usage.

The complete accuracy–efficiency comparison with all twelve baselines is provided in Section 5.4, Figure 5‾\underline{\text{Section 5.4, Figure 5}} of the revised manuscript.


We hope that our additional ablation studies and comparisons with recent baselines address your concerns and provide stronger evidence for the effectiveness and practical value of QUEST. Thank you again for your valuable feedback and consideration.

Thank you for the response. The new experiments address both of my main concerns. The comparison with TimePrism was especially helpful in showing where each method sits on the accuracy–efficiency Pareto frontier. I will raise my score and make the case for acceptance to the other reviewers and the AC during the discussion.

Thanks again for your insightful comments and positive feedback on our rebuttal. We truly appreciate the time and careful effort you put into reading our response and re-evaluating this work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.